AI Research Portfolio
A Scholarly Immersion at TERESOL, Islamabad
Institution
TERESOL, Islamabad
Supervisor
Javeria Ehsan
Duration
6 Weeks (Intensive)
Commitment
~10–12 hours/day
This intensive, research-focused internship served as an immersive experience in Artificial Intelligence. The six-week programme emphasised theoretical foundations, rigorous algorithmic implementation, and empirical evaluation across diverse learning paradigms. This portfolio showcases the progression from biomedical signal analysis and classical machine learning to deep learning, reinforcement learning, and real-time computer vision systems, demonstrating a sustained commitment to research excellence.
Foundations of Neural Signal Analysis
Research Review: EEG-Based Brain–Computer Interfaces (BCI)
A formal literature review was conducted on EEG-based BCI systems, specifically focusing on the application of machine learning and deep learning models for neural signal decoding. The analysis critically examined preprocessing complexities, noise sensitivity, inter-subject variability, and temporal dependencies inherent in EEG data.
A comparative study was performed between classical machine learning (e.g., SVMs) and deep learning architectures (CNNs, RNNs), establishing a robust biomedical AI perspective applied later in practical ECG projects.
Experimental Workflow and Data-Centric Thinking
All research projects adhered to a strict, consistent experimental workflow: data exploration, preprocessing, robust feature engineering, model training, evaluation, and diagnosis. This methodology emphasised understanding how preprocessing decisions directly influence model performance and generalisation capabilities.
Key research variables included the bias–variance tradeoff, rigorous dataset splitting strategies, and detailed error analysis, ensuring a rigorous, academic approach to model development across multiple datasets.
Technical Fluency: Numerical Computing and Visualization
Core technical fluency was established through extensive application of Python's scientific stack, leveraging powerful libraries for data analysis and visualisation.
Core Python Stack
Deep mastery of libraries including NumPy for numerical operations, Pandas for structured data manipulation, and Polars for high-performance computing.
Exploratory Visualization
Matplotlib and Seaborn were used not merely for reporting, but as essential analytical instruments for hypothesis testing, feature inspection, and rapid anomaly identification within datasets.
Correlation Heatmap
Histogram to show Skewness Box Plot to detect Outliers
Supervised Learning: Implementation and Comparative Analysis
A deep dive into supervised learning involved studying algorithms from both theoretical-mathematical and empirical perspectives, ensuring a comprehensive understanding of model mechanics and performance trade-offs.
Regression Models
Implementation and comparison of Linear and Logistic Regression, focusing on underlying assumptions and mathematical optimisation techniques.
Tree-Based Methods
Implementation of Decision Trees and Random Forests to understand non-linear decision boundaries and the application of ensemble methods for variance reduction and improved stability.
Proximity and Boundary
Exploration of K-Nearest Neighbors (KNN) and Support Vector Machines (SVMs), evaluating their robustness, generalisation capabilities, and interpretability across diverse datasets.
Visualisation of a Decision Tree
Linear Regression Prediction Line
Models were implemented both from scratch and via scikit-learn, with rigorous evaluation using appropriate metrics, focusing on generalisation performance.
Unsupervised Learning and Data Representation
Unsupervised techniques were employed to uncover latent structures and inherent relationships within complex, high-dimensional datasets, crucial for efficient learning.
K-Means Clustering
Applied K-Means clustering for pattern discovery and natural grouping identification within multi-dimensional feature spaces, assessing optimal cluster formation and cohesion.
Principal Component Analysis (PCA)
Utilised PCA for dimensionality reduction, focusing on variance preservation, data compression, and reducing feature redundancy in high-dimensional representations to improve model efficiency.
These techniques were critical in studying data geometry and improving downstream learning efficiency by providing meaningful, low-dimensional data representations.
K-Means Clustering with K=2 for visualising data separation.
Reinforcement Learning: Sequential Decision Making
Explored the dynamics of Reinforcement Learning (RL), focusing on agent interaction with complex environments and the mathematical framework for optimal policy derivation.
Reward Structures
Analysis of dense and sparse reward systems and their impact on efficient exploration and convergence speed in RL environments.
Policy Learning
Studied key algorithms such as Q-learning and SARSA, implementing them to derive optimal policies for sequential decision-making tasks.
Agent–Environment Loop
Designed a robust agent–environment interaction loop for standard Gymnasium environments, demonstrating control over state dynamics, reward signals, and termination conditions.
The study of RL highlights a capacity for non-static, adaptive thinking in artificial intelligence systems.
Deep Learning Architectures: Implementation and Analysis
Focused on building deep learning models from first principles, studying architectural intuition, feature extraction mechanisms, and design trade-offs across classic network types.
CNNs and RNNs
Implemented Convolutional Neural Networks (CNNs) for image tasks and Recurrent Neural Networks (RNNs)/LSTMs for sequential data, focusing on feature map creation and temporal dependencies.
Generative Models
Studied Autoencoders and Generative Adversarial Networks (GANs), with a focus on representation learning, data compression, and synthetic data generation capabilities.
Classical Architectures
Detailed analysis and scratch implementation of models such as AlexNet, ResNet, VGG, and GoogLeNet to understand parameter efficiency and hierarchical feature learning.
The emphasis was on architectural understanding, not just training accuracy, applied effectively to the FashionMNIST dataset and adversarial vulnerability analysis.
CNN Model Architecture
Le-Net-5 Performance Over FashionMNIST Dataset
LENET-5 Model Architecture
Sparse Autoencoder Performance Over FashionMNIST
Computer Vision and Real-Time Perception
Explored both classical and contemporary computer vision methods, focusing on developing efficient and robust real-time perception pipelines for practical deployment.
Classical Vision with OpenCV
Implemented foundational techniques using OpenCV, including advanced image preprocessing, edge detection algorithms (e.g., Canny), contour analysis, and template matching. This laid the groundwork for integrating traditional methods with learning-based models.
Real-Time Inference with MediaPipe
Developed real-time vision pipelines utilising MediaPipe for high-speed, accurate landmark detection (face and hand tracking). Emphasis was placed on optimising efficiency and robustness for resource-constrained environments.
Hand Landmark Detection
Object Detection on Videos
Template Matching Using Images
Capston Project: Healthcare AI - ECG Anomaly Detection
The culminating project focused on applying deep representation learning to a critical biomedical challenge: the unsupervised detection of cardiac anomalies from Electrocardiogram (ECG) data.
Signal Preprocessing
Rigorous filtering and normalisation of raw ECG signals to isolate key temporal features and mitigate noise interference.
Autoencoder Training
Trained sparse Autoencoders on large datasets of normal cardiac rhythms to establish a baseline model of healthy signal behaviour.
Anomaly Detection
Utilised reconstruction error as the primary metric for identifying anomalies. High error values indicate deviations from the normal pattern, suggesting potential pathologies.
This project demonstrated the capacity to integrate signal processing, deep learning, and robust evaluation methodologies within a healthcare context, directly addressing a problem of biomedical significance.
Visualisation of ECG signal characteristics and autoencoder reconstruction errors.
Internship Completion
Formal Conclusion and Certification
The rigorous research immersion concluded successfully at TERESOL in Islamabad, under the expert guidance of Supervisor Javeria Ehsan. The structured programme delivered comprehensive training across core AI methodologies, resulting in a formal certificate of completion and a robust academic portfolio.
